Best Strix Halo Mini PC for Local AI: Framework Desktop vs GMKtec EVO-X2 vs Beelink GTR9 Pro

Short answer: buy the Minisforum MS-S1 Max — $3,639 for 128GB/2TB with dual 10GbE, in stock on Minisforum's store (late-August shipping) as of early August 2026. ServeTheHome called it "the best Ryzen AI Max mini-PC yet," and right now it is the only one of the top boxes you can order online without a pre-sale wait or a store visit. The GMKtec EVO-X2 is the price floor at $1,999.99 list but every configuration showed unavailable; the Framework Desktop is $3,449 and out of stock; the Beelink GTR9 Pro repriced from ~$1,999 to $4,349 pre-sale and is off the board. All of them run the identical Ryzen AI Max+ 395 with 128GB, and all of them land in the same 30-45 tok/s band on gpt-oss-120b.

That last sentence is the key to this whole comparison, so let's say it plainly: there is one chip in every one of these boxes. Same 16 Zen 5 cores, same 40-CU Radeon 8060S, same 128GB of soldered LPDDR5X-8000. Nobody has a faster version. The vendors differ on ports, cooling budgets, power supplies, and — in 2026's memory shortage — wildly on price and stock. That makes this an unusually honest comparison to write: we don't have to pretend one badge produces magic tokens. We just have to tell you which box is the best-built, which is the best value, and which prices have quietly gone insane.

---

The Verdict Table {#verdict-table}

Bottom line first: identical silicon from $1,999.99 to $4,349 — and the two cheapest boxes were unbuyable when we checked, which makes the in-stock $3,639 Minisforum MS-S1 Max the de facto pick.

Prices below are from each vendor's own store, checked in the first week of August 2026. In this market they are perishable — treat them as a snapshot, not a promise.

Notice the pattern: the two cheapest listings are on boxes you cannot currently buy, and the best-reviewed box is the one actually shipping. That is the real state of this market, and pretending otherwise would make this page useless. The practical strategy in the memory-shortage era: if you need the machine this month, the MS-S1 Max at $3,639 is a fair price for the best-in-class box; if you can wait, set stock alerts on the EVO-X2 and Framework — anything at or under roughly $3,000 for 128GB is a win in 2026.

---

Same Chip in Every Box — So What Actually Differs? {#same-chip}

Every box here runs the AMD Ryzen AI Max+ 395: 16 Zen 5 cores, a 40-CU Radeon 8060S, and 128GB of LPDDR5X-8000 on a 256-bit bus (256GB/s). The differences are wattage, ports, power supply, and expansion — nothing else.

If you want the deep dive on the silicon itself — why unified memory changes the local-AI math, how much of the 128GB the GPU can actually address, quantization strategy — that's our Strix Halo / AI Max+ 395 guide. This page assumes the chip and asks only: which enclosure?

Four things genuinely separate the enclosures:

• Sustained power. GMKtec specs the EVO-X2 at 120W sustained, 140W peak. ServeTheHome measured "around 120 Watts" max during AI inference on the MS-S1 Max. Everyone is cooling the same ~120W reality; the difference is acoustics and headroom, not tokens.
• Networking. This is the biggest real spread: 2.5GbE (EVO-X2, HP Z2 Mini) vs 5GbE (Framework) vs dual 10GbE (MS-S1 Max, GTR9 Pro). If this box will serve models to your whole LAN, 10GbE is worth having.
• Power supply. Internal PSU (MS-S1 Max 320W, GTR9 Pro 230W, HP 300W, Framework FlexATX) means one cable and no brick. It's a small thing that matters every day the box sits on your desk.
• Expansion. The Framework Desktop and MS-S1 Max have a PCIe 4.0 x4 slot; everything else makes do with M.2 and USB4. None of them have enough PCIe for a real GPU, so treat the slot as a bonus (NIC, capture card), not a graphics upgrade path.

What does not differ: memory capacity, memory bandwidth, CPU, GPU. If a spec sheet or a listing implies one of these boxes is meaningfully faster than another, it's marketing.

---

What Performance to Expect {#performance}

Attributed numbers, one system, same silicon as every box here — ServeTheHome on AMD's Ryzen AI Halo Developer System (July 2026): gpt-oss-120b (MXFP4) ≈ 45 tok/s, Qwen3.6-35B (Q4) = 62 tok/s, dense Qwen3.6-27B ≈ 14 tok/s.

Those three numbers tell you almost everything about what this hardware is for. Verbatim, from the review:

Read the pattern, not the absolute numbers: mixture-of-experts models fly on Strix Halo; large dense models crawl. A 120B MoE only activates a few billion parameters per token, so 256GB/s of unified bandwidth is plenty. A 27B dense model drags every weight through memory for every token, and the same bandwidth becomes the ceiling. ServeTheHome adds a useful yardstick: about 11.6 tok/s sustained generates a million tokens a day. By that measure, a 45 tok/s gpt-oss-120b box is a serious little inference server, not a toy.

Between brands, expect single-digit-percent differences at most. ServeTheHome found the MS-S1 Max "beating all of the rest of our Ryzen AI Max+ 395 systems in both GPU workloads and multi-threaded workloads" — a real result, and also a small one. Runtime choices move numbers more than badges do: the same gpt-oss-120b ran at 31.41 tok/s in LM Studio on ServeTheHome's Beelink GTR9 Pro versus ~45 tok/s in their AMD dev-system testing at MXFP4. Same chip, same memory, same ~120W. Nobody buys their way out of physics with a better fan curve.

---

The Price Chaos: October 2025 vs Now {#price-chaos}

The same Beelink that ServeTheHome said "often sell for $1999 or less" in October 2025 lists at $4,349 on Beelink's own store today — up more than 100% in ten months. The memory shortage did this, and it did it unevenly.

128GB of LPDDR5X is exactly the commodity the memory supercycle is inflating, and each vendor has absorbed that differently:

Two honest observations. First, the cheap list prices and the buyable boxes are currently different boxes — GMKtec's store showed every EVO-X2 configuration "sold out or unavailable" at that attractive $1,999.99, and third-party marketplace listings during the shortage have run far above vendor list. Never pay a marketplace scalp for one of these; the whole point of Strix Halo is value per gigabyte. Second, of the boxes you can actually click "buy" on, one is fairly priced (MS-S1 Max, $3,639) and one has doubled (GTR9 Pro, $4,349). That's not a coincidence; it's what shortage pricing looks like.

---

Framework Desktop — the serviceable one {#framework-desktop}

$3,449 for the Max+ 395/128GB DIY config (out of stock in early August 2026). 4.5L case, internal FlexATX PSU, 5GbE, Wi-Fi 7, two M.2 2280 bays, and a PCIe x4 slot — plus a "192GB coming soon" tease on the product page.

The Framework Desktop is what happens when the repair-friendly laptop company builds a mini PC: a 4.5L case (96.8 × 205.5 × 226.1mm) around a standard Mini-ITX board with an internal FlexATX supply, 5GbE Ethernet, Wi-Fi 7, and two user-selectable front expansion-card slots carried over from their laptops. The board's PCIe x4 slot is real but not accessible in the default case — tinkerers will find a way; everyone else should treat it as latent.

The memory is soldered here like everywhere else (that's an LPDDR5X signal-integrity constraint, not a Framework choice — they were unusually upfront about it), so "serviceable" means storage, fans, board swaps, and a case you're allowed to open without a heat gun. The board-swap point is quietly the most interesting: Framework sells the mainboard as a part, and the company has publicly listed a 192GB follow-on. No other vendor on this page offers even a theoretical upgrade path to the Ryzen AI Max PRO 400 generation.

Buy it if: you keep hardware for five-plus years, you want the least-proprietary build, and it restocks near $3,449 the week you're buying. Skip it if: you need dual 10GbE (it has one 5GbE port) — the $3,639 MS-S1 Max adds that plus USB4 v2 for $190 more, and while Framework stays out of stock the question answers itself.

---

GMKtec EVO-X2 — the price floor, when it exists {#gmktec-evo-x2}

$1,999.99 on sale ($2,199.99 list) on GMKtec's own store — the cheapest 128GB Strix Halo path on paper — but every configuration showed "sold out or unavailable" when we checked. Single 2.5GbE port, 120W sustained/140W peak, dual M.2 up to 8TB each.

The EVO-X2 was one of the first Strix Halo minis (ServeTheHome reviewed it back in July 2025) and it is built to a price in visible ways: one 2.5GbE LAN port where the MS-S1 Max has two 10GbE, DisplayPort 1.4 alongside HDMI 2.1, and a chassis whose three-heat-pipe, dual-fan cooler is spec'd for the same ~120W everyone else sustains. One correction to something you'll see repeated in older roundups: GMKtec's current spec page lists no OCuLink port on the EVO-X2 — if an external-GPU path matters to you, plan around USB4 or pick a box with a PCIe slot.

None of that matters much at the right price, because the chip and the 128GB are the same as in boxes costing twice as much. A single-user desk box that runs gpt-oss-120b at the same ~45 tok/s as a $4,349 machine is a very easy thing to recommend at $1,999.99 — if you can catch a restock at that number.

Buy it if: it's in stock at or near list and you're one person at one desk who doesn't care about 10GbE. Skip it if: you're paying a marketplace premium (that erases its entire reason to exist), or the box will be a shared LAN inference server.

---

Beelink GTR9 Pro — great hardware, indefensible price {#beelink-gtr9-pro}

$4,349 on Beelink's store (list $4,699) for 128GB/2TB, on pre-sale with ~35-day shipping quoted. Dual 10GbE, two USB4 ports, internal 230W PSU — a genuinely nice box that now costs $350 more than AMD's own $3,999 developer system. At this price: no.

This one hurts, because the GTR9 Pro was the value story of late 2025. ServeTheHome's October 2025 review praised the Apple-inspired build, the Intel E610 dual 10Gbase-T networking, the internal PSU, and — pointedly — the pricing: "often sell for $1999 or less." The hardware hasn't changed. The sticker has: $4,349 as of early August 2026, a 100%+ increase in about ten months.

At $4,349 the comparison set stops being "other Strix Halo minis" and becomes everything: AMD's Ryzen AI Halo Developer System is $3,999 with first-party support and AMD's own preloaded software image, the MS-S1 Max matches the GTR9's dual 10GbE (and doubles its USB4 bandwidth) for $710 less — in stock — and if raw speed per dollar is the goal, a used 24GB GPU feeding our 24GB VRAM model picks demolishes every box on this page for anything that fits in 24GB. One more caution from the review cycle: ServeTheHome's GTR9 Pro coverage flagged Intel E610 NIC driver instability under sustained load, so budget a firmware/driver update before trusting the 10GbE for heavy transfers.

Buy it if: the price falls back under roughly $3,000, at which point it's an excellent machine again. Skip it if: the price is anywhere near today's. We don't say that about many products; identical silicon for $2,350 less on the same day's market makes this one easy.

---

The Other Three: MS-S1 Max, HP Z2 Mini G1a, and AMD's Own Box {#other-boxes}

The Minisforum MS-S1 Max ($3,639, in stock) is the box we'd actually buy; the HP Z2 Mini G1a is the corporate-procurement version; AMD's $3,999 Ryzen AI Halo Developer System is the zero-setup option.

Minisforum MS-S1 Max — the real pick. $3,639 for 128GB + 2TB (list $4,549), and — rare in this market — in stock, with late-August shipping quoted when we checked. ServeTheHome's January 2026 review titled it "the best Ryzen AI Max mini-PC yet" and the spec sheet explains why: dual 10GbE (Realtek RTL8127), two USB4 v2 ports at 80Gbps (the only box here with USB4 v2), a PCIe 4.0 x4 expansion slot, and a 320W internal PSU. The same review found it "beating all of the rest of our Ryzen AI Max+ 395 systems in both GPU workloads and multi-threaded workloads." Its flaws are small: HDMI-only video output that ServeTheHome noted "can be glitchy," and PCIe lanes spread thin across the extras. The buyable best-in-class box, $190 over an out-of-stock Framework, is an easy call.

HP Z2 Mini G1a — the workstation channel. HP's take uses the Ryzen AI Max PRO variants (up to the Max+ PRO 395 — same silicon, enterprise management bits) with up to 128GB, of which HP says up to 96GB is assignable to the GPU. You get a 300W internal PSU, dual NVMe with RAID, HP's modular Flex IO rear port options, and dual 2.5GbE — but no public sticker price; HP sells this through configure-and-quote. If your employer buys hardware through a VAR and wants warranty paper, this is the Strix Halo box that gets through procurement. If you're spending your own money, the consumer boxes above deliver the same tokens for less friction.

AMD Ryzen AI Halo Developer System — the first-party option. $3,999, 128GB, 2TB, 10GbE, sold first-party through AMD's pre-order program, shipping Debian with AMD's Ryzen AI Developer Center — one-click installs for vLLM, ComfyUI, and llama.cpp, plus pre-loaded playbooks (per ServeTheHome's review). It's the machine ServeTheHome's benchmark numbers above actually come from, and the box for people who value a supported software image over $1,000. Our memory-shortage buying map covers where it sits against NVIDIA's GB10 boxes if you're cross-shopping ecosystems.

If none of these fit the budget, our best mini PC for Ollama roundup covers the smaller-memory boxes where 7B-32B models live comfortably for a fraction of the price.

---

Setup: From Unboxing to First Token {#setup}

Whichever box you buy, the software path is identical, because the silicon is. Ollama is the ten-minute route; llama.cpp with Vulkan or ROCm is the tinkerer's route.

The one-liner install (from ollama.com's official Linux instructions), then the model that justifies the 128GB:

curl -fsSL https://ollama.com/install.sh | sh
ollama run gpt-oss:120b     # 65GB download — fits with ~60GB to spare
ollama run gpt-oss:20b      # 14GB — the warm-up act while the big one downloads

Those are the current tags on Ollama's official gpt-oss library page; the 120b weights ship MXFP4-quantized, which is exactly the format ServeTheHome's 45 tok/s figure was measured at.

If you want llama.cpp directly, the project's build docs give two AMD-friendly backends. Vulkan is the low-drama one on the Radeon 8060S:

# Vulkan backend (run vulkaninfo first to confirm your install)
cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release

# ROCm/HIP backend — find your GPU target first
rocminfo | grep gfx | head -1
cmake -S . -B build -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -- -j 16

Both command sets follow llama.cpp's official build documentation (omitting GPU_TARGETS compiles for the GPUs present, per the same docs). Two practical notes from the platform itself: give the GPU a large memory allocation in BIOS if your vendor exposes the option — HP's spec sheet caps GPU-assignable memory at 96GB of the 128GB, which is a good mental model for all of these boxes — and check our Strix Halo guide for the memory-allocation and quantization details that don't change per vendor.

---

Honest Limitations {#limitations}

A Strix Halo mini PC is a capacity play at ~256GB/s. It is not fast at dense models, it is not upgradeable, it is not CUDA, and right now it is not reliably in stock.

• Dense models are slow. ServeTheHome's 14 tok/s on dense Qwen3.6-27B is the number to internalize. If your daily driver is a big dense model rather than a MoE, a discrete GPU from our best GPUs for AI ranking will feel dramatically quicker for anything that fits in its VRAM.
• The memory is soldered. All of it. Everywhere. 128GB is what the box ships with and what it retires with. The only announced path past that is the 192GB Ryzen AI Max PRO 400 generation, which means a new board, not a stick.
• No CUDA. ROCm and Vulkan cover the mainstream stack (llama.cpp, Ollama, LM Studio) well, but if your workflow depends on CUDA-only tooling, no Strix Halo box fixes that — that's a different purchase.
• ~120W is the ceiling. Sustained power in the 120W class is why these boxes are silent and sippy, and also why no vendor can offer a "performance edition." What you see benchmarked is what everyone gets.
• Stock and price are the real spec sheet in 2026. The two best-value boxes were unbuyable at list the week we checked, and the buyable one had doubled in price. Set alerts, be patient, refuse scalps.

None of this is a reason to avoid the category — it's the reason to buy the cheapest well-built box rather than the fanciest one. The capacity is the product. Everything else is enclosure.

---

Which One Should You Buy? {#which-one}

Rules, in order: (1) cheapest acceptable box in stock wins; (2) under ~$3,000 for 128GB is a stock-alert win in 2026 — $3,639 for the best box, buyable today, is the realistic going rate; (3) dual 10GbE only matters if the box serves your LAN; (4) never pay marketplace scalp.

• Best overall, and in stock: Minisforum MS-S1 Max ($3,639). Best ports (USB4 v2, dual 10GbE, PCIe slot), biggest PSU, ServeTheHome's crown — and the one you can actually order this month. This is the buy.
• Best value, conditionally: GMKtec EVO-X2 ($1,999.99 list). Same tokens as everything else for the least money — strictly at or near list, in stock, for a single-desk user. Set the restock alert.
• Best for keeps: Framework Desktop ($3,449). The one designed to be opened, with a published 192GB future on the same platform — when it restocks.
• Best zero-effort: AMD Ryzen AI Halo Dev System ($3,999). First-party, with AMD's Developer Center image preloaded. You're paying $360 over the MS-S1 Max for a supported image.
• Wait on: Beelink GTR9 Pro. Fine machine, wrong year. Under ~$3,000 it re-enters the conversation instantly.

And a sanity check before any of them: if everything you plan to run fits in 24GB, a used discrete GPU is still the faster, cheaper answer — start from the hardware hub and the 24GB model picks before spending $3,000 on capacity you may not use.

---

Sources {#sources}

• ServeTheHome — AMD Ryzen AI Halo Developer System review (July 2026) — the gpt-oss-120b ≈45 tok/s, Qwen3.6-35B 62 tok/s, and dense-27B ≈14 tok/s figures
• ServeTheHome — Minisforum MS-S1 Max review (January 14, 2026) — "best Ryzen AI Max mini-PC yet," GPU-workload chart-topper, ~120W inference power
• ServeTheHome — Beelink GTR9 Pro review (October 11, 2025) — specs and the "often sell for $1999 or less" launch-era pricing
• ServeTheHome — GMKtec EVO-X2 review (July 11, 2025)
• Vendor stores, checked early August 2026: Framework, GMKtec, Beelink, Minisforum, HP Z2 Mini G1a
• Ollama gpt-oss library page and llama.cpp build documentation — setup commands

---

FAQ {#faq}